Research map / Graph learning / Graph self-supervised learning

DANCE: Dual Unbiased Expansion with Group-acquired Alignment for Out-of-distribution Graph Fairness Learning

ICML 2025 · Graph learning › Graph self-supervised learning · approach: Graph domain adaptation

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes DANCE, a GNN framework for fair node classification under graph distribution shifts that combines unbiased Mixup and adversarial graph expansion with group-acquired alignment and representation disentanglement.

Problem

Graph fairness learning assumes i.i.d. train and test nodes, but real graphs shift across communities/regions, causing sensitive group imbalance and conflicts between fairness and task performance under distribution shift.

Approach

Synthesizes unbiased virtual graphs in both graph space (confidence-guided minority-group Mixup with PPR/heat-kernel-based synthesized edges) and hidden space (fairness-aware adversarial feature generator), aligns representations with a group-wise contrastive objective that treats same-sensitive-label negatives specially, and disentangles sensitive from task representations.

Key results

• On Pokec-n, DANCE reaches 66.58% ACC and 72.88% ROC-AUC with delta DP 0.83% and delta EO 0.29%, the best overall rank among the compared methods.
• On Credit-Cs test communities, DANCE attains the best rank (1) with lowest delta EO in C2 (0.40) and C4 (0.36) and best ROC-AUC on C1 (72.18).
• Ablation: removing any of the four modules (Mixup, adversarial learning, group-acquired alignment, disentanglement) degrades DANCE on Credit-Cs, with representation disentanglement removal increasing delta DP and delta EO.

Measured effects reported in the paper

Surprising findings

• On several Credit-Cs communities, DANCE's accuracy is slightly lower than a variant (e.g., C2 ACC 78.4 vs. 80.58 for Variant 3), showing fairness and accuracy are not uniformly improved together.

Limitations

• Hyperparameters (lambda, beta, gamma, delta) need dataset-specific tuning; no single value is optimal across datasets.
• Evaluated only on three real-world graph datasets with binary sensitive attributes and binary labels.
• Theoretical convergence result relies on Lipschitz gradient assumptions and SGD, which do not directly cover the adversarial min-max training used in practice.

Experimental setup

Datasets: Credit-Cs Pokecs (Pokec-z train, Pokec-n test) Bail-Bs

Benchmarks: Credit-Cs Pokecs Bail-Bs

Metrics: ACC ROC-AUC delta DP delta EO combined rank score

Baselines: MLP GCN FairVGNN NIFTY EDITS CAF EERM SAGM RFR FatraGNN

Builds on: FatraGNN FairGNN Graph diffusion (PPR / heat kernel) Mixup EERM

Closest papers by meaning

• Fair Graph U-Net: A Fair Graph Learning Framework Integrating Group and Individual Awareness (AAAI 2025 · similarity 0.89)
• Chasing Fairness in Graphs: A GNN Architecture Perspective (AAAI 2024 · similarity 0.89)
• Disentangling, Amplifying, and Debiasing: Learning Disentangled Representations for Fair Graph Neural Networks (AAAI 2025 · similarity 0.89)
• Stable Fair Graph Representation Learning with Lipschitz Constraint (ICML 2025 · similarity 0.88)
• FairSIN: Achieving Fairness in Graph Neural Networks through Sensitive Information Neutralization (AAAI 2024 · similarity 0.87)
• Networked Inequality: Preferential Attachment Bias in Graph Neural Network Link Prediction (ICML 2024 · similarity 0.86)